Stefan Mittnik and Willi Semler

The Real Consequences of Financial Stress

Working Paper Number 07, 2013

Center for Quantitative Risk Analysis (CEQURA)
Department of Statistics
University of Munich

http://www.cequra.uni-muenchen.de
 The Real Consequences of Financial Stress∗

Stefan Mittnik† and Willi Semmler‡

First version: March 11, 2011; this versions: February 22, 2013

Abstract

We introduce a dynamic banking–macro model, which abstains from conventional mean–

reversion assumptions and in which—similar to Brunnermeier and Sannikov (2010)—adverse
asset–price movements and their impact on risk premia and credit spreads can induce instabil-
ities in the banking sector. To assess such phenomena empirically, we employ a multi–regime
vector autoregression (MRVAR) approach rather than conventional linear vector autoregres-
sions. We conduct bivariate empirical analyses, using country–specific financial–stress indices
and industrial production, for the U.S., the UK and the four large euro–area countries. Our
MRVAR–based impulse–response studies demonstrate that, compared to a linear specification,
response profiles are dependent on the current state of the economy as well as the sign and size
of shocks. Previous multi–regime–based studies, focusing solely on the regime–dependence of
responses, conclude that, during a high–stress period, stress–increasing shocks have more dra-
matic consequences for economic activity than during low stress. Conducting size–dependent
response analysis, we find that this holds only for small shocks and reverses when shocks be-
come sufficiently large to induce immediate regime switches. Our findings also suggest that,
in states of high financial stress, large negative shocks to financial–stress have sizeable positive
effects on real activity and support the idea of “unconventional” monetary policy measures in
cases of extreme financial stress.

Keywords: banking–sector instability, financial stress, monetary policy, nonlinear VAR, regime
dependence
JEL classifications: E2, E6, C13

∗
Previous versions, entitled “Estimating a Banking–Macro Model for Europe Using a Multi–regime VAR,”
were presented at the 19th Symposium of the Society for Nonlinear Dynamics and Econometrics, 2011; 4th
International Conference on Computational and Financial Econometrics, 2011; SEEK Workshop on Non-
Linear Economic Modelling: Theory and Applications, 2012; and workshops at Humboldt University Berlin
and Giessen University, Germany, as well as the Centre Cournot, Paris, BI Business School, Oslo, Monetary
Authority of Hong Kong, and the Catholic University, Milano. We thank the participants for valuable
comments. We are especially grateful for communications with Markus Brunnermeier and to James Ramsey,
Timo Teräsvirta and two anonymous referees for constructive comments, which helped to greatly improve
the paper. Part of the research was done while SM was Theodor Heuss Visiting Professor at the New School
for Social Research and while WS was Fulbright Visiting Professor at the Vienna University of Economics
and Business and visiting the Center of Economic Risk at Humboldt University Berlin, being supported by
the German Research Foundation via the CRC 649 Economic Risk. We thank the hosting institutions for
their excellent hospitality and the funding organizations for their generous support. WS would also like to
thank the Centre for European Economic Research (ZEW) for financial support.
†
Dept. of Statistics and Center for Quantitative Risk Analysis, Ludwig–Maximilians–Universität
München, Akademiestr. 1/I, 80799 Munich, Germany.
‡
Dept. of Economics, New School for Social Research, 79 Fifth Avenue, New York, NY 10003, USA
1 Introduction

The banking system played a central role in the 2007-08 financial meltdown in the U.S.
This phenomenon, as Reinhard and Rogoff (2009) and Gorton (2009, 2010) underscore, is all
too common in financial crises and tends exacerbate and amplify the economic downturn—
whatever may have caused it. Gorton (2010) points out that whereas in the past loan losses
and bank runs have been the conventional mechanisms by which crises where triggered,
more recently, banking crises seem to be strongly related to adverse shocks in asset–value
losses and financial stress. In the aftermath of this recent “great recession,” a number of
studies have investigated the effects of financial stress on economic output. He et al. (2008),
Brunnermeier (2009), Adrian et al. (2010), Davig and Hakkio (2010), and Hubrich and
Tetlow (2011) conduct empirical analyses for the U.S. economy. Monnin and Jokipii (2010)
consider several OECD countries; Mallik and Sousa (2011) and Hollo et al. (2012) look at
the euro area; and van Roye (2012) focuses on Germany.
Theoretical analyses of the recent meltdown have largely used used the financial accelerator
to model the linkage between the financial and the real sector. This strategy is not without
problems. Brunnermeier and Sannikov (2010, 2012) point out that the destabilizing dynamics
are not adequately captured in such a setting. Furthermore, the financial–accelerator theory
has been mainly applied to firms and households, so far. Bernanke, Gertler and Gilchrist
(1999) show that financial markets can have amplifying effects. In the DSGE tradition,
however, there are only locally magnifying effect, namely, through collaterals. Collateral
values rise at a high level of economic activity, making credit available and cheap; and the
reverse happens at a low level of economic activity.
From a technical viewpoint, DSGE–type models suffer from the fact that the are often solved
via linearization about a unique and stable steady state, and amplifying effects occur only
with respect to deviations from the steady state. Moreover, they typically do not track
debt dynamics.1 Departures from the steady state are eventually mean reverting. Although
the economy might accelerate, ultimately it will revert back to the steady state. Related
empirical analyses are often conducted by means of linear vetor autoregressions (VARs) as
in Gilchrist et al. (2009, 2011), Christensen and Dib (2008), and Del Negro et al. (2010).
As the recent meltdown has, however, demonstrated, shocks to banks seem to be destabilizing
rather than mean reverting.2 Important empirical papers in this context are Brunnermeier
1
Empirically, the debt–to–asset ratio is predicted to fall in booms and to rise in recessions (cf. Gilchrist
et al., 2009). Yet, as Geanakoplos (2010) mentions, the empirical measure is distorted through the way the
debt–to–asset ratio is measured.
2
Students of the great depression maintain that credit and the banking sector had quite strong destabi-

1
(2009) and Brunnermeier and Pederson (2009), who show that financial intermediaries3 often
have to liquidate their capital, when asset prices fall and margin requirements. This forces
financial intermediaries to take a hair cut and to further de-lever, which depresses asset
prices even further and, thus, reinforces the downward spiral.
Models attempting to capture such mechanisms often stress that falling asset prices, possibly
accompanied by a fire sale of assets, have external effects on the financial sector. The
possibility of a downward spiral, then, comes from interconnectedness, interlinkages and
contagion. Investigations along these lines began with Greenwald and Stiglitz (1996) and,
more recently, continued with Geanakoplos and Farmer (2009), Adrian et al. (2010), Gorton
(2010), Geanakoplos (2010), and Brunnermeier and Sannikov (2010, 2011). These studies
argue that such dynamics create an endogenously generated jump in risk, which is usually
triggered by large changes in asset–price movements.
Much of the recent research states that this process primarily works through the banks’
balance sheets. In the first instance, banks may have loan losses that may arise from defaults
of firms or households, the foreign sector, or from sovereign debt. On the other hand,
large shocks to asset prices and financial stress affect the asset and liability sides of banks’
balance sheets, reducing the availability of credit. As the financial stress rises, so will risk
premia, repo rates, TED spreads, and other credit spreads. These spillover effects to other
intermediaries (as well as to firms and households) create what Brunnermeier and Sannikov
(2010, 2012) refer to as endogenous risk.
In this paper, we investigate theoretically and empirically the question of how financial
stress and economic output interact. Specifically, we study theoretically how destabilizing
mechanisms might work in a model that captures linkages between the banking sector and
the real sector. Looking at the U.S. and five EU counties, namely, the UK and the four
largest euro–zone economies Germany, France, Italy and Spain, we examine to what extent
such linkages can be detected and quantified empirically, using country–specific financial
stress indices constructed by the IMF and discussed Cardarelli et al. 2009.
Rather than, as Brunnermeier and Sannikov (2010, 2011), emphasizing the role of asset prices
and asset–price volatility in downward destabilization, our approach focuses on movements
in risk premia and credit spreads. This is motivated by theoretical and empirical studies
lizing effects. Minsky (1976, 1982) and Kindleberger and Aliber (2005), for example, regard the credit sector
as a significantly amplifying force. The latter view the instability of credit and Minsky the way financing
becomes de-linked from collaterals as contributing factors to a downward spiral, once large real or financial
shocks occur. This important tradition captures many aspects of the banking–macro link.
3
This may include Gorton (2010) calls the shadow banking system, i.e., investment firms, brokers and
money–market dealers, that has grown sizeably in the last 15 to 20 years.

2
which show that factors, such as large asset price falls, rising volatility, higher risk premia,
and a run into liquidity, are highly correlated with movements in discount rates.4 Yet, as
we will show, triggering of downward instability also depends on the constraints imposed on
the banking industry, such as on the growth of capital assets (through borrowing) and on
payouts, affecting banks’ risk taking, equity formation and leveraging. Higher payouts, for
instance, may induce more risk taking and risk transfer and, thus, generate higher aggregate
risk and risk premia to be paid by all. We will explore the dynamic behavior for different
constraints on banks’ decision variables.
Our model is in the spirit of Brunnermeier and Sannikov (2010, 2011), however, we introduce
model variants, where the amplifying mechanisms might be stronger or weaker. In a version
with strong endogenous amplification of risk (greater credit spreads) and less constraints
on insiders’ payouts and on growth rates of capital assets (leveraging), there is a greater
likelihood for instabilities to occur. On the other hand, in a version with little amplification
of credit spreads and stricter constraints on payouts and growth rates of capital assets, insta-
bility is less likely to arise. By making this distinction, our analysis offers more differentiated
results as, for example, compared to Brunnermeier and Sannikov (2010, 2011).
Standard linear econometric models, such as vector autoregressions (VARs), cannot capture
the rich dynamic behavior of our model, as it implies that responses to financial shocks
are state dependent. In a state of high financial stress with substantial credit spreads, a
positive, financial stress–increasing shock may have different effects than a negative, stress–
reducing, shock. The same shocks occurring in a stage of low financial stress, low credit
spreads and, possibly, high economic activity, may have, again, very different implications.
To allow for such phenomena, our empirical analysis employs nonlinear, multi–regime VARs
(MRVARs). Such models enable us to capture the complex dynamics and to analyze how
the economies under investigation respond at different economic states to different types
of shocks. To estimate the interactions between financial stress and economic output, we
conduct bivariate analyses using industrial production and country–specific Financial Stress
Indices constructed by the IMF.
The paper is organized as follows. Section 2 develops a model that reflects the features
discussed above. Section 3 describes the numerical solution of the two model variants pro-
posed, using dynamic programming techniques. Section 4 discusses the empirical analysis by
describing the data, briefly introducing the MRVAR approach and presenting the empirical
findings from VAR and MRVAR response analyses. The results for the U.S. are discussed in
4
The different factors driving discount rates—and, through that, asset prices—are extensively discussed
in Cochrane (2011).

3
 Table 1: The basic balance sheet of a bank

Assets Liabilities
pt kt dt
nt =pt kt -dt
total assets total liabilities

some detail, those for the five EU countries are summarized. Section 5 concludes.

2 The Banking Model and Its Dynamics

In our model we follow Brunnermeier and Sannikov (2010, 2011) (in short, BS), but in-
troduce decision constraints that can lead to varying degrees with which financial crises
have amplifying effects. We also differ in the way we solve the model, namely, by applying
dynamic–programming techniques.
To model the banking sector, we look at the basic balance sheet of a bank (see Table 1). On
the left hand side, we have the assets, pt kt , valued at current asset prices, pt , and, on the
right hand side, total debt, dt , and the resulting net worth, nt = pt kt −dt .5 The balance sheet
comprises the variables that form a bank’s decision problem. We assume banks maximize
objective function ˆ ∞
V (k, d) = max E e−ρt U (ct )dt, (1)
ct ,gt 0

with the state variables capital stock and debt evolving according to

dkt = (ϕ(it /kt ) − δ)kt dt + σkt dZt (2)

and
ddt = (rt dt − (akt − it − ct ))dt, (3)

respectively, with rt denoting the interest rate paid on debt. System (1)–(3) defines the
model which underlies our analysis.
Along the lines of BS, we presume that there are payouts, bonus payments for executives,
which can be viewed to serve the decision makers’ consumption stream, ct ,6 which is dis-
5
BS (2010) state that equity might be split into inside equity, α(pt kt − dt ), and outside equity, (1 −
α)(pt kt − dt ). In the following, we will not make distinction.
6
Semmler and Bernard (2012), compiling the bonus payments of the six largest U.S. investment banks,
report that bonus payments, as a percent of revenues, went up from roughly 10% in 2000 to about 35% in

4
counted at rate ρ. As BS, we take log utility to represent preferences. Next to the consump-
tion stream, investment, i.e., gt = it /kt in (2), represents the second decision variable. As
implied by (3), investments are permitted to be financed by borrowing.
In (2), assets of the financial intermediaries will be increased by investments, it /kt , where
function ϕ(it /kt ) includes adjustment cost and is concave in the argument, and δ represents
the depreciation rate of assets.7 The actual gross capital of the bank increases at the rate
it /kt . The last term in (2) is the diffusion term in form of a Brownian motion, with σ being
constant.
According to (3), debt evolves at a rate determined by the excess spending of investment and
consumption over capital income, with the latter being linear, i.e., akt . While investments
will increase the stock of assets, a high rate of asset purchases will increase debt. The interest
rate to be paid on debt, rt , includes a risk premium, which is tends to reflect weak balance
sheets and, thus, financial strain of a bank. Below, the interest rate will be endogenously
determined, reflecting state or time depending risk premia.
Our setup amounts to a standard model of wealth management, as commonly used to study
wealth management of financial institutions (see He and Krishnamurthy, 2008). If we replace,
in (3), the constant income for a unit of wealth, a, by a weighted average of risky and risk–free
returns on a wealth fund, then, our system has similarities with those found in the wealth–
management literature (cf. Semmler et al., 2009). There, however, an explicit equation for
the evolution of debt of the financial intermediary, as in (3), is typically missing. This is an
innovative contribution of BS (2010, 2011) and others.8
BS (2010, 2011) specify the growth rate of asset prices as a geometric Brownian motion,
namely,
dpt = µt pt dt + σt pt dZt , (4)

which is characterized by time varying volatility, σt . This can, for example, arise during
rapid sales of assets and create endogenous risk. Instead of using asset price movements and
varying volatility, we employ time– or state–dependent risk premia and credit spreads to
capture such effects.
2007. In its attempts to reforming the regulation of financial markets, the EU plans to restrict the cash
component of bonus payment to 20% of total bonus payments, with the remaining portion only to be paid
out in subsequent years via common stocks. In our model, we ignore such details in the bonus system.
7
In a recent version, BS (2011) specify a model with capital assets that could be banking capital or real
capital. The model is equivalent of there is perfect substitution among them.
8
See, for example, Hall (2010), who also includes an equation for the evolution of debt. BS (2010) argue
that before 2007 financial intermediaries have been encouraged to take on more risk by being able to transfer
of risk to outside investors—for example, via CDOs. As a consequence, financial intermediaries had built up
debt and, thus, default risk.

5
We model debt via the debt–asset ratio, dt /kt ,9 and, as BS (2010), define ω = −(dt /kt ).
Taking log and time derivative, we can write asset–accumulation and debt dynamics by
modifying objective function (1), namely,10
ˆ ∞
V (ωt ) = max E e−ρt U (c˜t )dt, (5)
c˜tt ,gt 0

such that
dωt = [(gt − rt − δ + σ 2 )ωt + a − τ (gt )]dt − c˜t +σωt dZt , (6)

with the consumption–to–wealth ratio, c˜t = ct /kT , being an new control variable.11
Expression τ (gt ), in (6), represents a convex adjustment cost, which is affecting the size of
borrowing to achieve growth rate gt . This is modeled by following the capital–adjustment–
cost literature. However, only the growth of wealth, gt , enters the evolution of assets, kt .12

3 Specifications, Solution Method and Numerical Re-

sults

Clearly, and as stressed by BS (2010, 2011), the dynamics of models of the type (1)–(3)
or (5)–(6) cannot not be studied using the usual linearization techniques. First or even
second–order Taylor approximations to solve for local dynamics will not adequately capture
the global properties of the model. We, therefore, employ the dynamic–programming (DP)
approach put forth in Gruene and Semmler (2004) to study the compact version (5)–(6). To
do so, we explore the out–of–steady–state dynamics using a coarse grid for a larger region of
the state space and gradually refine the grid for smaller regions. DP provides information
on the truly global dynamics in a larger region of the state space with little loss in accuracy
(see Becker et al., 2007).13
In our model versions, the debt–to–asset ratio is the state variable, and the decision variables
are the growth rate of assets and payouts for consumption. Moreover, as in Semmler et al.
9
Note that we use stocks of assets and debt, in contrast to Geanokoplos (2010), who uses flows as leverage
measure with leveraging being highly positively correlated with booms.
10
For a similar approach, see also Hall (2010).
11
A derivation of a dynamic equation in the stochastic case, using Ito‘s lemma, is given in BS (2010, 2011).
Note, in (6), we treat δ as a constant.
12
The other expressions in (6) are straightforward derivations from the negative of the growth rate of the
debt–asset ratio as stated above.
13
A short summary of the solution method of a dynamic decision problem, such as (5)-(6), is given in
Ernst and Semmler (2010).

6
(2009), time needs to be introduced as a further state variable. For the decision variables,
we will impose both loose or tight constraints and, as specified below, will let credit spreads
be state– and time–dependent.

3.1 State-Dependent Credit Spreads and Soft Decision Constraints

We first study a model variant with state-dependent credit spreads and soft decision con-
straints. One issue in the financial–accelerator theory has been that, relative to empirical
evidence, the state–dependent risk premium is too small. To better match the data, Nolan
et al. (2009) and Gilchrist et al. (2011) have introduced shocks on risk premia that exhibit
strong persistence and increase the financial accelerator effect.
To specify two regimes of low and high risk premia—and, thus, two credit–spread regimes—
one can define credit spreads as a nonlinear function of leveraging, ωt . To do this, in con-
tinuous form, we define credit spreads via

rt = κ arctan(ωt ). (7)

The arctan(·) function, with κ > 0, avoids the extreme instabilities arising in the asset–price–
leveraging dynamics as might be implied in Geanokoplos (2010) and BS (2010, 2011). In
our DP algorithm, we restrict ourselves to the simplest case where banks pay risk premium
according to (7).14 Banks face credit spreads that rise as the leverage increases. Thus,
decisions on consumption and asset growth indirectly depend on state–dependent credit
spreads that financial intermediaries are exposed to (cf. Gruene and Semmler, 2004).
Payments15 can be tightly of loosely constrained. BS (2010, p. 17) already conjecture that,
when bonus payouts are less constrained, “the system is relatively stable near its ‘steady
state’ ... but becomes unstable below the steady state ...”
In the first version of our model, we impose soft decision constraints. When defining the
payout, c˜t , we restrict the choice to the interval cmin
˜ < c˜t < cmax
˜ . Thus, payouts are always
16
positive, but constrained. For the case of less constrained payouts, we set 0.01 < c˜t < 0.3.
Moreover, we constrain the growth of assets to −0.1 < gt < 0.1.17
14
Specification (7) makes credit spreads endogenous, but not without bounds. The arctan function for the
risk premium has a lower limit, the risk–free rate, as well as an upper limit.
15
Constraints on the growth rate of assets (or certain types of assets) acquired by banks, are harder to
impose, unless there are borrowing constraints introduced, as Geanokoplos (2010) seems to suggest.
16
Note that we also could allow for dividend payments, as the model can encompass dividend payments.
17
We do not elaborate on the origin of these constraints, i.e., whether they are imposed by the regulator
or self–imposed. We only explore the effects of different constraining rules.

7
 dZ
0.75

0.45

0.15

-0.15

-0.45

-0.75 ω
-2.083 -1.583 -1.083 -0.583 -0.083 0.417

Figure 1: State-dependent credit spread, large payouts, and small domain of attraction

BS (2010, p. 32) state that allowing the debt–to–asset ratio rise too much—as a result of
intermediaries’ incentives to take on risk for the sake of high (short term) payouts—may
induce downturns. In their view, downturns in financial, product and labor markets and
higher asset–price volatility result from the fact that externalities are not taken fully into
account.
We now present some numerical solutions for the model, reporting first the results for state–
dependent credit spread and loosely restricted decisions. We parametrize the model as
follows: a = 0.25, α = 0.3, σ = 0.008, and ρ = 0.03. Note that, given the soft constraints
on asset growth and consumption payouts, both together can easily surpass the gross return
from assets, akt , quickly generating excessive leverage.
Figure 1 depicts how the debt–asset ratio, dt /kt , shown on the horizontal axis, reacts to
stochastic shocks, dZt , shown on the vertical axis, with the standard deviation set to σ =
0.008. The path of ω varies in the state space and there is no unidirectional vector field; i.e.,
the path of ωt is not a straight line. For low values of the debt–asset ratio, the numerical–
solution path in Figure 1 exhibits an unstable steady state at ω ∗ = −0.85. Since zero is a
steady state, there is only a small region of attraction.
In other words, small initial leveraging will eventually lead to low or zero debt–asset ra-
tios, because the credit spread is low. A large shock, driving the debt–asset ratio above
ω ∗ = −0.85, causes both the leveraging and the credit spread to rise, leading to a vicious

8
cycle: higher leverage creates higher credit spreads and higher credit spreads induce higher
leveraging.18 On the other hand, there is a small domain of attraction: if the financial
intermediary starts with low leveraging and low credit spread, both may decrease further.
It is the debt–asset ratio—and the corresponding credit spread—together with the loose
restriction of the decision variables that lead to the amplifying behavior. Thus, as predicted
in BS (2010), under given conditions, a leveraging ratio beyond a certain threshold is likely
to be dynamically unstable. 19

3.2 Time–depending Credit Spreads and Stricter Decision Con-

straints

We now turn to the case of a small variation in risk premia and credit spreads and tighter
restrictions on the decision variables. As a proxy for time varying in credit spreads, we
extract low–frequency components from FSI data.20 By doing so, we smooth extremes and
tend obtain less erratic dynamics.21 In our DP procedure, we would then have one additional
state variable, such as
dx = 1dt, (8)

a time index that captures the time varying risk premia. The low–frequency components in
the credit spread is indexed on x, representing time in the DP algorithm. It can be computed
and included in our numerical procedure, so that, in (3) or (6),

rt = rx (x). (9)

Formally, the stochastic, dynamic decision problem now has two decision variables and three
state variables, the leverage ratio, ωt , the time index, xt , and the stochastic shock term, dZt ,
with rt being time varying.
BS (2010, p. 32) state that limiting payouts should be welfare improving, arguing that “...
a regulator can improve social welfare by a policy that limits bonus payments within the
18
Here, we do not address the question at what leverage ratio bankruptcy might occur. This depends on
the distance to default, which is defined in the KMV model by the standardized distance of the bank’s asset
value to its debt level. This is issue is pursued in Gruene and Semmler (2005).
19
Since the shape of the value function for this case is similar to the next case, it will be discussed in sect.
3.2.
20
Credit spreads, such as the BAA/AAA spread and the BAA/T–Bill spread, are important FSI compo-
nents. Spread measures that have been widely used in previous studies (cf. Gilchrist et al., 2009).
21
Details of computing the time–varying risk premia are given in the appendix.

9
financial sector. Specifically, suppose that experts are not allowed to pay themselves as long
as financial intermediaries are not sufficiently capitalized.” This type of regulatory effort
would keep the banking sector sufficiently capitalized and, thus, make it more stable.22
Regulation could require payouts to be cut, if the ratio of net worth to total assets falls
below a certain threshold. In other words, if ω = −(dt /kt ) 5 $, then payouts are lowered or
even set to zero. Payout rules could also be designed so that management has an incentive
to reduce leverage, by allowing high payouts only when leverage is lower.23 This might
be considered as a penalty on risk taking and high leveraging.24 With the constraints on
decision variables imposed, the dynamics of the debt–wealth ratio are likely to change.25
For the case of more tightly constrained payouts, we restrict (negative and positive) growth
rates of intermediaries’ asset purchases to the interval −0.1 < gt < 0.08 and also constrain
the consumption–to–capital ratio to 0.01 < c˜t <0.18. Both choices are optimal, but they are
constraints. Given these constraints, growth of capital assets and consumption payouts is
less likely to exceed the return on capital, akt .26
Figure 2 shows that, compared to the looser specification, the domain of attraction is in-
creased, with the steady state, now roughly at ω ∗ = −3.9, being a repeller: with lower asset
growth and lower payouts, the debt–to–asset ratio will go to zero. The domain of attraction
of the zero debt–to–asset ratio is considerable larger. The potential for large externalities,
financial stress and meltdowns are reduced. An initial debt–to–asset ratio to the right the
threshold, ω ∗ = −3.9, given stricter constraints and time varying risk premia, will always
preserve stability. And a high debt–to–asset ratio can be stabilized, if risk premia and
payouts are small.27
22
A similar view is present in the Geneva Report (2009, Section 6.2 and the Basel III guidelines.
23
This has been suggested, for example, by the Financial Stability Board (2009, p. 3): “For significant
financial institutions, the size of the variable compensation pool ... should take into account the full range
of current and potential risks ..."
24
BS (2010) argue higher bonus payouts may encourage more risk taking and risk transfer, generating
eventually higher aggregate risk and greater risk premia.
25
One could consider further modifications that take into account the availability of funds from households
and from capital inflows, see Caballero and Krishnamurthy (2009). There was also a particular inflow of funds
from Central Banks, as in the U.S. in the years 2008 and 2010, when the Fed conducted its unconventional
quantitative–easing policies, i.e., buying bad and rapidly declining assets from financial intermediaries. The
ECB provision of a three year low interest rate liquidity for EU banks starting in December 2011 and the
recent purchase of sovereign bonds is are similar cases. All this has a mitigating effect on the unstable forces
generated by the banking system. On the other hand, the precautionary motives of households (and firms)
and a “flight to high quality assets” lead to a reduction of financial funds for financial intermediaries.
26
U.S. regulatory effort have considered to lower the limit on the asset–to–equity ratio from 33% to 12%,
as Basel III intends. Our second model specification is more in line with stricter constraints on the asset–
to–equity ratio.
27
In Figure 2, the shocks, dZt , are drawn from the range −0.1 < dZt < 0.1, and we have −10 < ωt < 0.

10
 dZ
0.030

0.012

0.006

-0.006

-0.018

-0.030

t
33.333

0.000 ω
-10.000 -8.000 -6.000 -4.000 -2.000 0.000

Figure 2: Time–varying credit spread, small payouts and large domain of attraction

Figure 3, showing the corresponding value function, reveals that total welfare (for the finan-
cial intermediaries) is rising with lower debt–to–asset ratios. The rise of the value function
to the right of the steady state is reasonable, since the welfare from lower credit spreads
should be higher.28
In both model specifications, the risk premium, and thus the credit spread, are state and
time dependent. It is the state dependency of the credit spreads and the softer constraints
that are likely to trigger a vicious cycle. In the more restrictive setting, we observe a much
larger domain of the zero attractor of the debt–to–asset ratio. Thus, larger shocks are less
consequential than in the less restrictive case. The downward instability depends not only on
the level of financial stress—and, thus, the size of the risk premia—but also on the constraints
of the decision variables.
For the third dimension, the time index xt , we specify 0 < t < 100. The graph, because we have movements
in a three–dimensional space (reflected by the additional time axis), shows the trajectory for a fixed t.
28
The value function corresponding to Figure 1 looks similar and also increases with the debt–to–asset
ratio falling, i.e., ω moving to the right of −0.85. Note that the shape of the value function is roughly the
same as shown in BS (2010) in their Figure 7, though we have negative values on the vertical axis, as we use
log c˜t rather than c˜t in the preferences.

11
 V

dZ

Figure 3: Value function for time-varying credit spread, small payouts, and large domain of
attraction

The two specification could be interpreted as two different types of banking systems, which
are (in the first case) more or (in the second case) less prone to experience instabilities. As
a consequence, the reaction of the two systems to financial shocks may be quite different.
Analogously, the success of policy measures, designed to stabilize an ailing financial sector
and to avoid or revert negative real consequences, will depend on the specific characteristics
of the banking system.
In the following section, we report on the results of an empirical multi–country study that
allows for the possibility of complex financial–sector dynamics and the interactions between
financial stress and real activity. To address these questions, we employ multi–regime VARs
and response–analysis techniques.

4 Empirical MRVAR Analysis

In order to investigate the empirical empirical evidence for the possibility of banking sectors
being characterized by smaller and larger domains of attraction, we study the stability
properties for six countries, the U.S. and the five largest EU economies. Specifically, we
will examine whether the data support the single–regime, linear VAR specification for the
six countries, or whether they favor the nonlinear MRVAR hypothesis, which implies the

12
existence of varying domains of attraction and state–dependent dynamics. Before doing so,
we discuss our variable choices for measuring financial stress and output.

4.1 Empirical Measures for Real Activity and Financial Stress

Our measure for real activity is dictated by the data frequency chosen for the empirical
analysis. To investigate short–, medium– and longer–term effects, our study is based on
monthly data, so that the growth rate of industrial production (IP) becomes a “natural”
proxy for real activity. The fact that, in some cases, the relative size of the industrial sector
differs substantially from country to country may result in country–specific heterogeneity in
our empirical finding.
The discussion in the previous section may suggest the leverage ratio to be a trigger–variable
for banking instability. However, it can be questioned whether the ratio of net worth to
capital assets, or its reciprocal, the degree of leveraging, ω, represents an appropriate measure
of financial stress, as it is greatly affected by the way markets valuate assets and liabilities.
The valuation of assets is typically the result of their income–stream estimates and presumed
discount rates. Similarly, the valuation of liabilities, such as bonds or short– and long–term
loans, is strongly affected by their respective risk premia.29 Moreover, various forms of
credit constraints—measured, for example, by the Fed index of changes in credit standards
to determine the ease and tightness of obtaining credit as well as default premia, credit
spreads and short term liquidity—represent stress factors for financial intermediaries.
As all these factors potentially affect the credit demand and supply of financial intermedi-
aries, a more comprehensive measure than simply leverage appears to be called for. There-
fore, in our empirical analysis we employ the IMF’s financial stress index (FSI), which are
consistently constructed for a large number of advanced and developing countries (see Ca-
radelli et al., 2011).30 .
The FSI considers three major sources and measures of instability, namely, (i) a bank–related
component: a 12–month rolling banking beta and the TED spread; (ii) a securities–related
component: a corporate bond yield spread, an inverted term spread, and a monthly stock
returns (measured as declines), six–month rolling monthly squared stock returns, and (iii)
an exchange rate component: a six–month rolling monthly squared change in real exchange
rates.
29
This is implicit in Merton’s risk structure of interest rates, Merton (1974).
30
A variety financial stress indicators have been developed in recent years. Kliesen et al. (2012) analyze
and compare eleven alternative measures for the U.S. that are constructed from different subsets of altogether
109 fundamental variables. See also Illing and Lui (2006) and Hatzius et al. (2010)

13
 GER ESP FRA

FSI
15 log IP 180 15 180 15

10 160 10 160 10

5 5 5
140 140 140
0 0 0 130
120 120 120
−5 −5 −5 110
100 100 100
−10 −10 −10 90
1980 1990 2000 2010 1980 1990 2000 2010 1980 1990 2000 2010

ITA GBR USA

20 20 200

15 15 180

10 160 10 10 160
150
5 5 5
140 140 140
0 0 130 0
120 120 120
−5 −5 110 −5
100 100 100
−10 −10 90 −10
1980 1990 2000 2010 1980 1990 2000 2010 1980 1990 2000 2010

Figure 4: Time series of financial stress (dark line, left scale) and scaled industrial production
(light line, right scale)

Figure 4 plots the FSI and (scaled) IP series, starting December 1980, for the four large euro
area countries Germany, Spain, France and Italy as well as Great Britain and the U.S. The
graphs reveal an inverse relationship between FSI and IP. Especially when financial stress
exhibits sudden large positive spikes, IP tends to drop. More generally, we observe that, in
general, low financial stress and a well–performing real–sector coincide.31 Given the apparent
relationship between financial stress and economic activity, we also expect a strong linkage
between net worth of financial intermediaries and economic activity, as financial stress affects
their balance sheets.
To further assess the relationship between industrial production and financial stress, we con-
duct (linear) Granger–causality tests for the six countries. The results reject the hypothesis
31
This coincides with the empirical study in Gorton (2010) that insolvency rates of financial institutions
increase in bad times.

14
 Table 2: p-values from Granger–causality tests between IP growth and FSI

Country IP → FSI FSI → IP

GER 0.0001 0.4986
ESP 0.8074 0.0778
FRA 0.0128 0.0709
ITA 0.0184 0.0707
GBR 0.0036 0.1574
USA 0.0000 0.0006

“IP growth does not Granger–cause FSI” (first column in Table 2) in five out of six cases
with p–values below 0.02, with Spain being the sole exception (p = 0.8074). For the reverse
hypothesis, i.e., “FSI do not Granger–cause IP growth” (second column in Table 2) all p–
values exceed 0.05, except that for the U.S. (p = 0.0006). However, with values ranging from
0.07 to 0.08, the results for the southern EU countries France, Italy and Spain are somewhat
more borderline.
The results on Granger causality should be regarded here as a descriptive tool that summa-
rizes correlational, i.e., linearly approximated, relationships. In view of the strong evidence
against the linearity assumptions reported below, the results have to be interpreted with
caution.

4.2 MRVAR Approach

Linear VAR models have been frequently used to study the financial accelerator.32 Such
“single–regime VAR” analyses may be suitable for studying the local behavior of between
variables affected by infinitesimal disturbances. They are, however, not expected to provide
meaningful insights into consequences of large shocks, nor do they allow for state–dependency
in shock responses or asymmetric effects to positive and negative shocks. Similarly, so–called
“corridor stability,” as discussed in the earlier literature on Keynesian macro dynamics (cf.
Dimand, 2005; Bruno and Dimand, 2009) and also referred to in the context of financial–
market regulation (cf. Schinasi, 2005)—i.e., situations where small shocks have only small
effects but larger shocks could have large, destabilizing effects—cannot be handled in con-
ventional VAR specifications.
32
Christensen and Dib (2008) estimate financial accelerator for the macroeconomy with a linear VAR.
For the application of the financial accelerator to study financial intermediaries in a linear VAR setting, see
Hakkio and Keeton (2009) and Adrian et al. (2010).

15
To address such questions empirically, we require a more general modeling framework that
can accommodate varying dynamic patterns. The multi–regime vector autoregression (MR-
VAR)33 approach we adopt permits us to detect the presence of multiple regimes and to
investigate regime–dependence in the responses to shocks to the system. Shocks might oc-
cur during a regime with great instability, as, for example, in the case of a high leverage ratio
(or high stress) discussed in Section 3 (see Figure 1) the effects will be larger as compared
to a regime with a large domain of attraction and low leverage (and low stress), see Figure
2. As a consequence, responses to positive and negative shocks may not be sign–symmetric
and different shock sizes may lead to nonproportional response patterns.
To allow for such phenomena, we generalize the linear, constant–parameter VAR specifi-
cation in the “mildest” form and adopt a piecewise linear VAR, such as Markov–switching
autoregressions (Hamilton, 1989) or threshold autoregressions (Tong, 1978, 1983). A charac-
teristic of Markov–switching autoregressions is that the states are unobservable and defined
by some combination of the endogenous variables, so that they do not necessarily have an
obvious interpretation. Also, a given observation cannot directly be associated with any
particular regime. Only conditional probabilistic assignments are possible via statistical
inference based on past information.
For our purposes, namely state–dependent response analysis, states are associated with
specific levels of financial stress. MRVAR models in the form of threshold autoregression
models of Tong (1978, 1983) or, in a vector setting, of multivariate threshold autoregressions
(Tsay, 1998) are obvious candidates. In contrast to Markov–switching autoregressions or
standard multivariate threshold autoregressions, our approach assumes that we can, based
on some observable variable, define upfront a set of interpretable regimes, which are not a
result of some estimation procedure, but rather motivated by the objective of the empirical
analysis.
The threshold–based MRVAR approach is a simple and parsimonious strategy for nonpara-
metric function estimation and for modeling multi-equilibria settings (Hansen, 2000). It
is particularly suitable for our analysis, where we are interested in evaluating the poten-
tial effectiveness of policy measures for specific, observable states of the economy, in our
case high–stress and low–stress states, which we can explicitly specify by defining threshold
variables and, possibly, threshold levels.
33
For an MRVAR–modeling application see also Mittnik and Semmler (2012) and Ernst et al. (2010).

16
The MRVAR specification we adopted below is given by
pi
X
yt = ci + Aij yt−j + εit , εit ∼ (0, Σi ), if τi−1 < rt−d ≤ τi , for i = 1, . . . , M, (10)
j=1

where rt−d , d > 0, is the value of the threshold variable observed at time t − d; and regimes
are defined by the threshold levels −∞ = τ0 < τ1 < · · · < τM = ∞. In the following analysis
we estimate a two–regime VAR, with the FSI as threshold variable and the threshold level
being estimated by minimizing the corrected AIC (AICc) model–selection criterion.
Response analysis for linear VAR models is straightforward. Point estimates and asymp-
totic distributions of shock response can be derived analytically from the estimated VAR
parameters (cf. Mittnik and Zadrozny, 1993). In nonlinear settings, this is, in general,
not possible, and one typically has to resort to Monte Carlo simulations. Following Koop
et al. (1996), the so–called generalized impulse responses, which depend on the overall
state, zt , the type of shock, vt , and the response horizon, h, are defined by GIRh (zt , vt ) =
E (yt+h | zt , ut + vt ) − E (yt+h | zt , ut ), where the overall state, zt , reflects the relevant infor-
mation set. For a Markov–switching VAR process, zt comprises information about the past
realizations of yt and the states; for an MRVAR, only information about past realizations
yt−1 , · · · , yt−pmax , with pmax = max(p1 , . . . , pM ), is required.
To understand the regime–specific differences in the dynamic characteristics, within–regime
response analysis, as in Ehrmann et al. (2003), are helpful. This assumes that the process
remains in a given regime for the next, say, h periods and amounts to conducting standard
VAR–type response analysis for each regime estimated. Within–regime response analysis
can also be viewed as a study of the local dynamic behavior, as shocks need to be sufficiently
small for the system not to jump to another regime for the next h periods.

4.3 Estimation

We conduct bivariate analyses for the six countries shown in Figure 4, using monthly
IP growth and IMF’s stress index data from 1981 to 2012.34 Specifically, we estimate
standard VAR and MRVAR models for the IP–growth rate and FSI levels, defining yt =
(100∆ log IPt , F SIt )0 . For the MRVARs, we specify FSI as the regime–defining threshold
variable, as is also done in Hollo et al. (2012), setting the delay parameter in (10) to d = 1.
This allows us to fit the dynamics according to an economy’s state of financial stress and to
34
We use seasonally–adjusted industrial–production data from the OECD (2012); the FSI data were pro-
vided from the IMF (2011). See Table 3 for the country–specific sample periods.

17
assess, for example, the effectiveness of policy measures intended to boost real activity by
lowering financial stress.
Model selection for the MRVAR specification (10) can be done using the AIC

M
X
AIC (M, p1 , . . . , pM ) = AICi , (11)
i=1

where  
n+3
AICi = Ti ln |Σ̂i | + 2n npi + ; (12)
2
and M is the number of regimes; pi is the autoregressive order of regime i; Ti reflects the
number of observations associated with regime i; Σ̂i is the estimated residual covariance
matrix for regime i; and n denotes the number of variables in vector yt . Formulation (11)
differs from that in Chan et al. (2004) in that we account for possible heterogeneity in the
constant terms, ci , and residual covariance, Σi , across regimes.35
Alternatively, Wong and Li (1998) suggest the use of the corrected AIC (in short, AICc),
which tends to give more parsimonious specifications and is obtained by modifying (11) via

XM  
2ki (ki + 1)
AICc (M, p1 , . . . , pM ) = AICi + , (13)
i=1
Ti − ki − 1

with ki = pi n2 + n(n + 3)/2. Based on univariate simulation experiments, Wong and Li

(1998) strongly recommend the AICc over both the AIC and the BIC—a recommendation
which we follow here.
The model specifications we obtain for the six countries are reported in Table 3. A look
at the AICc values suggests that the MRVAR model clearly dominates the linear VAR
specification, with AICc differences ranging from 13.6, for Great Britain, up to 48.7, for
Germany. Although judging by such difference does not represent a formal test, the widely
accepted “rule of thumb” for the AIC (see Burnham and Anderson, 1998) states that there is
“essentially no support” for a model when the difference to the minimum AIC is larger than
10.
The estimated—from a goodness–of–fit point of view—optimal threshold levels, τ̂ , defining
high– and low–stress regimes and, thus, their domains of attraction, vary substantially from
35
When employing (11) to discriminate between an MRVAR and a standard VAR specification (i.e., a
single–regime MRVAR), we need to include the n parameters in the intercept vector, c, and the n(n + 1)/2
parameters in the residual covariance matrix for an equivalent parameter count.

18
 Table 3: Specifications of VAR and MRVAR models
Germany Spain France Italy UK U.S.
VAR
p 3 5 3 4 2 4
AICc 793.1 662.1 454.2 626.8 305.2 -31.1
T 377 376 374 376 377 377
MRVAR
τ̂ -1.949 -2.343 -0.402 1.313 2.519 3.062
plo 2 1 4 4 2 4
phi 3 5 3 1 1 1
AICc 744.4 630.4 410.0 600.7 291.6 -70.1
Tlo 132 75 192 279 317 327
Thi 245 301 182 97 60 50
Sample 01/1981 01/1981 04/1981 01/1981 01/1981 01/1981
period 05/2012 04/2012 05/2012 04/2012 05/2012 05/2012

country to country. The values for the U.S., the UK and Italy are positive, so that only
more extreme, above–average stress periods are classified as high–stress states. For Germany
and Spain, we obtain negative threshold estimates, associating most of the sample with with
high–stress and only FSI–values below about -2 are treated as low–stress situations. Note
that, for each country, the VAR lag lengths and the (maximum) MRVAR lag lengths are
identical. The sole exception is France, where pV AR = 3 and, for the MRVAR, plo = 4 and
phi = 3.

4.4 Response Analysis for the U.S.

We now discuss the empirical results for the U.S. in some detail and summarize those for
the other countries in Section 4.5. To assess the effects of the linear versus the nonlinear
model specification, we first look at the estimates of the cumulative unit–shock responses
for the VAR model and then regime–specific responses for the MRVAR model. The latter
give an indication for how the system responds to small shocks that do not induce migration
to another regime. Then, we move to the more realistic and—for example, for policy–
making—more relevant analysis by investigating overall system responses by allowing for
regime migration, looking first at unit–shock responses and then at the size–dependency.
For all responses analyses, we assume that a shock to IP simultaneously affects the FSI,
whereas IP reacts with a one–period delay to an FSI shock.

19
 VAR Resp of IP to IP Shock VAR Resp of IP to FSI Shock
3 0.2

0
2.5

−0.2
USA 2
−0.4

1.5
−0.6

1 −0.8
0 10 20 30 40 0 10 20 30 40

VAR Resp of FSI to IP Shock VAR Resp of FSI to FSI Shock

0 10

8
−2

6
−4
4

−6
2

−8 0
0 10 20 30 40 0 10 20 30 40

Figure 5: Standard VAR responses for the USA

4.4.1 Linear VAR Responses

The cumulative responses due to a unit shock implied by the estimated linear VAR model
are shown in Figure 5 over a horizon of 36 month.36 The results suggest that, for the U.S., a
positive one–standard–deviation stress shock has an increasingly negative effect on IP, which
settles at about -0.65% after three years. The cumulative response of IP to a unit shock
in IP itself settles at 2.6%; and the stress index responds negatively to a positive IP shock
(-7.2%), whereas a positive FSI shock lets the index increase (8.6%).

4.4.2 Within–regime MRVAR Responses

Next, we explore the MRVAR within–regime response behavior. Clearly, the assumption
to stay within a particular regime for an extended period is not very realistic as shocks
and regime dynamics may induce regime migration. Also, by looking at the within–regime
dynamics, we solely focus on the regime–specific autoregressive parameters and ignore the
level effects resulting from differences in the regime intercepts. Any differences in the regimes’
intercepts will induce additional variation in the overall dynamics when the process switches
36
In the discussion of the general results from the response analysis we focus solely on the responses’
point–estimates. We will consider interval estimates when we discuss the results specific to the question
under investigation.

20
 High FSI: Resp of IP to IP shock High FSI: Resp of IP to FSI shock
2 0
USA

1.5 −0.2

1 −0.4
0 10 20 30 40 0 10 20 30 40
High FSI: Resp of FSI to IP shock High FSI: Resp of FSI to FSI shock
10 4

0 2

−10 0
0 10 20 30 40 0 10 20 30 40
Low FSI: Resp of IP to IP shock Low FSI: Resp of IP to FSI shock
2 0

1.5
−0.2
1

0.5 −0.4
0 10 20 30 40 0 10 20 30 40
Low FSI: Resp of FSI to IP shock Low FSI: Resp of FSI to FSI shock
0 4

3
−0.5
2

−1 1
0 10 20 30 40 0 10 20 30 40

Figure 6: Within–regime responses from MRVAR for USA in high–FSI (top half) and low–
FSI regimes (bottom half)

between regimes. However, a regime–specific response analysis is useful as it helps to better

understand the short–term dynamics associated with the estimated regimes.
The within–regime MRVAR responses are presented in MRVAR in Figure 6. Here, the cumu-
lative responses to unit–shocks implied for each of the two MRVAR regimes are somewhat
different. Within the high–FSI regime (upper half in Figure 6) a positive one–standard–
deviation stress shock lowers IP by -0.30%, which is reached after one year, and amounts
to about half of the reduction implied by the linear VAR. Compared to the VAR analysis,
at high–FSI, IP responds less positively to a shock to IP itself, settling at 1.8% rather than
2.6; the MRVAR stress response to a positive IP shock is less negative (-5.4%); and the cu-
mulative MRVAR response to a positive FSI shock is lower than the VAR–implied estimate
(3.5% versus 8.6%).
Within the low–FSI regime (lower half in Figure 6) a stress shock has a weak negative impact
on IP (-0.26%). With 1.7%, the IP response to a positive shock to IP is similar to that during
high stress. The stress index responds much less in states of high stress (-0.94 versus -5.4)
to a shock in IP; and the cumulative response to an FSI shock is somewhat stronger than

21
 High FSI: Resp of IP to positive FSI Shock High FSI: Resp of IP to negative FSI Shock
0.5 2.5

0 2
−0.5
1.5
USA −1
1
−1.5

−2 0.5

−2.5 0
0 10 20 30 40 0 10 20 30 40

Low FSI: Resp of IP to positive FSI Shock Low FSI: Resp of IP to negative FSI Shock
0.4 1.2

0.2 1

0 0.8

−0.2 0.6

−0.4 0.4

−0.6 0.2

−0.8 0

−1 −0.2
0 10 20 30 40 0 10 20 30 40

Figure 7: Cumulative MRVAR responses of U.S. IP to positive (left panel) and negative FSI
shocks (right panel) in high– (upper panel) and low–stress states (lower panel)

obtained for the high–stress regime (3.9 versus 3.5).

4.4.3 MRVAR–system Response Analysis to Unit–shocks

Given our objective, namely the evaluation of the impact of financial sector stress on eco-
nomic growth, measured in terms of IP growth, subsequent analysis will focus on the response
of IP to shocks to FSI. In order to investigate the economy’s overall growth effects due to
FSI shocks we, first, simulate generalized cumulative response functions to unit shocks, i.e.,
response functions wthat take the possibility regime migration into account. We do this for
specific states at which the shock is assumed to occur. The two specific states we select are
the sample averages observed for each of the two regimes, as they can be viewed as typical
representatives for low–and high–stress states of the economy. The low–FSI regime–mean
is ȳlo = (0.2372, −1.0127)0 , and that for high FSI ȳhi = (−0.2686, 5.5704)0 . From these we
simulate both a positive and a negative unit–shock to the stress index. The mean cumulative
IP responses together with one–standard deviation confidence bands are shown in Figure 7.
One striking difference between the system responses and the linear VAR and within–regime
MRVAR responses is that the former have not fully settled after after 36 months, indicating
that the system as a whole has more inertia and that shocks have a longer–lived impact than
linear response analysis suggests.

22
Table 4: MRVAR–implied cumulative IP responses after 36 months due to positive and
negative unit–shocks during high– and low–stress states. Note: the corresponding (state–
independent) linear VAR estimates are ±0.64.

IP response to:
Positive FSI shock Negative FSI shock
High stress -1.10 1.27
Low stress -0.47 0.34

The estimated unit–shock responses up to a horizon of 36 months, summarized in Table

4, strongly suggest that the impact of an FSI shock on IP varies with financial stress. A
positive unit–shock in the average high–stress state (top left plot in Figure 7) causes U.S.
IP to drop by about -1.1% within three years, after which zero is included in the confidence
band. The same shock applied in the average low–stress state (bottom left plot), results
in an IP contraction that is less dramatic (-0.43%) over a three–year horizon, where the
confidence band starts to include zero. Thus, in a state where financial stress is already
high, an additional unit–increase curbs IP more severely, namely about two and a half times
as much, than during calm periods. If, on the other hand, the FSI–unit–shock is negative,
we obtain pretty much the reverse results. In absolute terms, a negative FSI unit–shock
during high stress (top right plot) has a stronger impact than a positive one stress (1.27%
vs. -1.10%); with 0.34%, the weakest impact comes from a negative unit–FSI–shock when
the stress level is already low. Thus, at least for one–standard–deviation shocks occurring
during average high– and low–stress states, U.S. IP responses are strongly state dependent
and, to a lesser extent, sign–asymmetric.

4.4.4 MRVAR–system Responses to Different Shock Sizes

Next, we investigate to what extent the size of the shock to financial stress matters. In
addition to simply assuming a unit shock to the stress index, we simulate the cumulative IP
responses to FSI shocks with different sizes. Specifically, we impose positive and negative
shocks from one through four standard deviations.37 It turns out that the IP responses
differ dramatically as the magnitude of the shocks varies. Figure 8 compares the response
profiles scaled by the size of the shock;38 and Table 5 reports the point estimates of the
37
A shock–size of four standard deviations appears rather large. However, the FSI for the U.S. peaked in
during the 2008–09 stress period at 17.64, which corresponds to 5.44 standard deviations. Based on the FSI
data prior to that period, the peak corresponds to 6.99 standard deviations.
38
Note that all responses shown in Figure 8 turn out to increase monotonically, so that shocks of size one
have the weakest and four–unit shocks the strongest response. This also holds for Figure 9 through 13.

23
 High FSI: IP resp to pos FSI shock High FSI: IP Resp to neg FSI shock

15 Shock +1 15
Shock +2
Shock +3
10 Shock +4 10
USA

5 5
Shock −1
Shock −2
0 0 Shock −3
Shock −4
0 10 20 30 0 10 20 30

Low FSI: IP resp to pos FSI shock Low FSI: IP resp to neg FSI shock

15 Shock +1 15
Shock +2
Shock +3
10 Shock +4 10

5 5
Shock −1
Shock −2
0 0 Shock −3
Shock −4
0 10 20 30 0 10 20 30

Figure 8: Cumulative MRVAR responses of U.S. IP to negative (top) and positive (bottom)
stress shocks in high– (left) and low–stress states (right)

scaled responses after month 36.

If the economy is financially distressed, stress reduction is an effective measure to induce
growth. A reduction of one–unit induces a 1.27% cumulative increase in IP growth after 36
months (see upper half, last column in Table 5). A negative two– instead of a one–unit shock
is relatively more effective, resulting in a 3.05% increase per shock unit; and, compared to
this, a three–unit reduction has a substantially stronger implication with a per–unit boost of
12.71. This increase in relative impact is due to the fact that a stress reduction of this size
has very high probability of moving the economy quickly into a low–stress state. We do not
observe such a tremendous jump when the shock size grows even further, as the probability
of moving right away into the low–stress regime cannot anymore increased by that much.
The potency of negative shock is considerably less in low– than in high–stress periods (lower
half, last column, Table 5): one– and two–unit shock have about a quarter, and three– and
four–unit shock have less than a thenth of the impact compared to high–stress states.
Positive FSI shocks lower IP in both high– and low–stress states (first column, Table 5).

24
Table 5: MRVAR–implied relative cumulative IP responses after 36 months due to positive
and negative shocks of different sizes during high– and low–stress states. Note: all response
plots are scaled by the shock size and, thus, represent responses relative to unit–shock
responses; the corresponding scaled (state– and size–independent) VAR–implied estimates
are all ±0.64.

Shock Shock sign

size Positive Negative
High stress 1 -1.10 1.27
2 -1.99 3.05
3 -2.80 12.71
4 -3.67 14.83
Low stress 1 -0.47 0.34
2 -1.02 0.72
3 -1.74 1.02
4 -2.84 1.31

However, the negative impact is larger during high stress, ranging from -1.10% to -3.67% for
shock sizes one through four, compared to only -0.47% to -2.84%. Thus, during financially
calm periods, IP reacts more (less) strongly to positive (negative) FSI shocks—which is just
the reverse from what we find for a financially distressed U.S. economy.

4.5 Results for Selected EU Countries

We now summarize the empirical findings for the five EU countries, the four largest euro
zone economies German, France, Italy, and Spain plus, not belonging to the euro area, the
UK.39 To do so, the discussion is based on the graphical presentations—analogous to Figure
8—of the IP–growth responses due to FSI shocks of different sizes and signs, when applied
at average high–and low–stress states. For each country, the stress–states are defined by
their respective threshold estimates.
The response profiles for the five countries, shown in Figures 9–13, and that of the U.S.
have, to a large extent, common patterns. But there are also a number of distinct features.
Especially Spain, the country under investigation that has been most troubled during the
euro crisis, exhibits a response pattern that is quite different.
Below, we first summarize our findings by looking at the response patters for the four possible
39
The respective VAR and MRVAR estimation results are summarized in Table 3, reporting for each
country the lag order, AICc values, estimated threshold values, and the number of observations associated
with high– and low–stress regimes.

25
 High FSI: IP resp to pos FSI shock High FSI: IP Resp to neg FSI shock
10 10
Shock +1
Shock +2
5 Shock +3 5
Shock +4
GER

0 0

Shock −1
−5 −5 Shock −2
Shock −3
Shock −4
0 10 20 30 0 10 20 30

Low FSI: IP resp to pos FSI shock Low FSI: IP resp to neg FSI shock
10 10
Shock +1
Shock +2
5 Shock +3 5
Shock +4

0 0

Shock −1
−5 −5 Shock −2
Shock −3
Shock −4
0 10 20 30 0 10 20 30

Figure 9: Cumulative MRVAR responses of German IP to negative (top) and positive (bot-
tom) stress shocks in high– (left) and low–stress states (right)

combinations of shock signs and states of stress, letting shock size, again, assume values ±1
through ±4.
Low stress and negative FSI shocks: The responses for this case are shown in the bottom,
right graphs in Figures 9–13. For four of the five the EU countries, we find that, in state
of low stress, IP remains virtually unaffected by FSI–reducing shocks. For shock sizes -1
though -4, the cumulative IP responses are practically zero. I.e., the countries’ real sectors
react even less than is the case for the U.S., which exhibits small positive growth effects.
In case of Spain, the response plots appear to be rather different, having opposite signs.
Looking, however, at the magnitudes, the responses are close to zero–responses.
High stress and positive FSI shocks (top, left graphs): The response profiles from positive FSI
shocks at financially distressed states of the economy are, except for Spain, for all countries
very similar. We estimate weak negative IP effects, which, given the scaling of the responses
by shock–size, increase in relative terms as the magnitude of the shock increase. In other
words, larger shocks are more effective per unit of shock. The response magnitudes vary
from country to country, with the U.S. exhibiting the largest impacts, followed by Germany,

26
 High FSI: IP resp to pos FSI shock High FSI: IP Resp to neg FSI shock
6 6
Shock +1
Shock +2
4 Shock +3 4
Shock +4
2 2
FRA

0 0
Shock −1
−2 −2 Shock −2
Shock −3
−4 −4 Shock −4
0 10 20 30 0 10 20 30

Low FSI: IP resp to pos FSI shock Low FSI: IP resp to neg FSI shock
6 6
Shock +1
Shock +2
4 Shock +3 4
Shock +4
2 2

0 0
Shock −1
−2 −2 Shock −2
Shock −3
−4 −4 Shock −4
0 10 20 30 0 10 20 30

Figure 10: Cumulative MRVAR responses of French IP to negative (top) and positive (bot-
tom) stress shocks in high– (left) and low–stress states (right)

the UK, Italy, and France. In the case of Spain, the responses have, again, the opposite sign
but are extremely small.
High stress and negative FSI shocks (top, right graphs): Excepting Spain, we find for all
countries a similar, positive IP response patterns for stress–reducing shocks in high–stress
states. As in the case of the U.S., for three out of the four countries (France, Italy and the
UK), the relative response–impact jumps when moving from the two– to a three–unit shock.
For Germany, the jump occurs when the shock size increases from three to four. This can
be due to Germany’s low FSI–threshold estimate, so that, in the average high–stress state,
more sizeable, negative shocks are required to move the economy into the low–stress regime.
The response sizes vary from country to country. The U.S., Italy, the UK, and—disregarding
the difference in the low–stress activation level—Germany display similarly large IP increases,
when the low–stress regime does not tend to kick in, exceeding those of by France quite a
bit. In case of large, negative four–unit shocks, U.S. IP reacts most, followed by Italy and
Germany with about two third of the impact, and the UK and France with about only 40%
of the impact estimated for the U.S.

27
 High FSI: IP resp to pos FSI shock High FSI: IP Resp to neg FSI shock
10 10
Shock +1
Shock +2
5 Shock +3 5
Shock +4
ITA

0 0

Shock −1
−5 −5 Shock −2
Shock −3
Shock −4
−10 −10
0 10 20 30 0 10 20 30

Low FSI: IP resp to pos FSI shock Low FSI: IP resp to neg FSI shock
10 10
Shock +1
Shock +2
5 Shock +3 5
Shock +4

0 0

Shock −1
−5 −5 Shock −2
Shock −3
Shock −4
−10 −10
0 10 20 30 0 10 20 30

Figure 11: Cumulative MRVAR responses of Italian IP to negative (top) and positive (bot-
tom) stress shocks in high– (left) and low–stress states (right)

Low stress and positive FSI shocks (bottom, right graphs): For all six countries, positive FSI
shocks during low stress behave just the opposite of what we find in the reverse case with
low stress and negative FSI shocks. The effects are, however, less pronounced than in the
latter case. Also, the shock sizes, triggering large impacts, tend to differ.

4.6 Relation to Similar Response Studies

Regime–dependent response analyses, investigating the effects of a shock to financial stress

economic activity using multi–regime approaches, have been conducted for the U.S. (Hakkio
and Keeton, 2009; and Hubrich and Tetlow, 2011), for Germany (van Roye, 2012) and the
euro area (Hollo et al. 2012). They come to the conclusion that a positive, one–unit stress
shock in high–stress periods induces a more dramatic reduction in economic activity than
during low–stress. These studies are, however, confined to the analysis of (positive) unit
impulse shocks and do not investigate the sign– and size–dependence of the response profile
as done here. For all countries (ignoring Spain), we also find that unit–shock responses have

28
 High FSI: IP resp to pos FSI shock High FSI: IP Resp to neg FSI shock

Shock +1
0.5 Shock +2 0.5
Shock +3
Shock +4
0 0
ESP

−0.5 −0.5
Shock −1
Shock −2
−1 −1
Shock −3
Shock −4
0 10 20 30 0 10 20 30

Low FSI: IP resp to pos FSI shock Low FSI: IP resp to neg FSI shock

Shock +1
0.5 Shock +2 0.5
Shock +3
Shock +4
0 0

−0.5 −0.5
Shock −1
Shock −2
−1 −1
Shock −3
Shock −4
0 10 20 30 0 10 20 30

Figure 12: Cumulative MRVAR responses of Spanish IP to negative (top) and positive
(bottom) stress shocks in high– (left) and low–stress states (right)

a more dramatic impact in high–stress than in low–stress regimes (see the left column in
Table 5 and the unit–responses in the left panels in Figures 8–11 and 13).
Our results go beyond those in the studies mentioned above, as they imply that, as the shock
size increases, this phenomenon will reverse. Large positive shocks will ultimately cause the
financial system to move from the low–stress to the high–stress regime, causing IP cuts that
are larger than experienced when the economy was already in a high–stress state.
In case of negative, i.e., stress reducing shocks, we obtain analogous results. Large stress
reductions in a high–stress state can induce a switch to a low–stress regime and, accordingly,
lead to large IP increases. On the other hand, a stress decrease during a calm period will—
in a two–regime setting—not result in a regime switch and, thus, leave IP more or less
unaffected (see the right column in Table 5 and the responses in the right panels in Figures
8–11 and 13).

29
 High FSI: IP resp to pos FSI shock High FSI: IP Resp to neg FSI shock

Shock +1
6 6
Shock +2
4 Shock +3 4
Shock +4
2 2
GBR

0 0
−2 −2 Shock −1
−4 −4 Shock −2
Shock −3
−6 −6 Shock −4
0 10 20 30 0 10 20 30

Low FSI: IP resp to pos FSI shock Low FSI: IP resp to neg FSI shock

Shock +1
6 6
Shock +2
4 Shock +3 4
Shock +4
2 2
0 0
−2 −2 Shock −1
−4 −4 Shock −2
Shock −3
−6 −6 Shock −4
0 10 20 30 0 10 20 30

Figure 13: Cumulative MRVAR responses of UK IP to negative (top) and positive (bottom)
stress shocks in high– (left) and low–stress states (right)

4.7 Response Analysis and Policy Implications

The response characteristics resulting from our MRVAR analysis have general implications
for policy action. First, the effectiveness of monetary policy measures increases with the size
of the measures; and the effectiveness increases over–proportionally.40 Thus, to calm extreme
financial stress, extreme measures are called for in order to support real–sector economic
activity. The Fed’s recent drastic and “unconventional” quantitative–easing policies, designed
to reduce the overall systemic risk, is in line with our findings, as only sufficiently large
measures are likely to move a financially stressed economy into calmer waters. Second, the
empirical results strongly suggest that the timing of policy actions affecting financial stress
is very influential on their success. Stress–reducing measures applied during calm periods
will have little real impact, relative to applying the same actions in a high–stress state.
Our findings are compatible with recent studies arguing that unconventional monetary policy
40
However, as the—short-run and long-run—cost of policy measures will also increase disproportionately
as their size increases, an appropriate balance needs to be found.

30
is needed in a financially distressed economy, reflected by a sharp rise in credit spreads,
which—rather than asset–price volatility—constitute the dominant component of the stress
index (cf. Curdia and Woodford, 2009). Thus, not only a decrease in interest rates, but a
reduction in credit spreads and financial stress as a whole is required to induce significant
expansionary effects on real economic growth.
More generally, the empirical findings suggests that the stronger the global position of an
economy in terms of output and share of world trade, the more autonomous are the financial
stress effects directly impacting the economy. Moreover, in these economies, a large stress
reduction in a high–stress period has a relatively more sizeable improvement on output. Yet
the smaller the economies are, the more they are subjected to external dynamics and the
spill over effects, which seem to significantly affect the consequences of domestic financial–
sector shocks. Moreover, as in the case of France, a larger public sector seems to act as a
buffer against (positive and negative) stress shocks—possibly, as compared to Germany, at
the expense of long–term growth.
The Spanish economy reacts very differently to shocks to the financial sector. Before the
crisis, its economic growth paralleled that of Germany. But that was mainly driven by a
booming housing sector, which seems to have insulated Spain from external effects. After
the—yet to be overcome–collapse of the real–estate sector, the economy has not recovered,
as it has the U.S. and Germany, and longer adjustment time may be needed to regain
diversification and competitiveness.

5 Conclusions

Most severe economic crises have led to a meltdown of the banking sector, and—regardless
of the crisis’ origin—the banking sector has typically aggravated the economic downturn. To
better understand the linkages between economic activity and financial stress, first, we have
developed a theoretical model describing the dynamics of the banking sector’s instability
when being exposed to shocks in asset price, credit spread and financial stress in general.
Then, in a six–country study, using nonlinear, multi–regime VARs and monthly data on
industrial production and a financial–stress index provided by the IMF, we have examined
real–sector and financial–sector interactions.
Our theoretical model implies that unfavorable credit spreads, high leverage ratios and ex-
cessive (bonus) payouts play an important role for the deterioration of the banks’ balance
sheets and the stability of the banking sector—or the lack thereof. In the presence of large

31
credit spreads, high leveraging and large payouts, sever adverse shocks can induce insta-
bilities, whereas, for smaller shocks, the system exhibits mean reversion. In contrast to
previous studies of the financial accelerator, which postulate a locally amplifying but glob-
ally stable and mean reverting system, our model admits downward instability as suggested
in Brunnermeier and Sannikov (2010).
The empirical findings reveal that the effects of financial–sector shocks are state–dependent
and that, in addition, the relative effectiveness of the shocks depends on the sign and size
of the shocks. Although there is some heterogeneity between countries in the sense that
IP responses to stress–shocks are larger in the bigger economies, such as Germany and the
U.S., there are quite common features: Large positive financial–sector shocks during states
of low stress have less of a contractionary effect than in a high stress regime; and large stress
reductions lead to stronger expansionary effects in high– than in low–stress regimes. This
holds for five out of the six countries studied, with Spain being the sole exception.
In terms of policy implications, our results strongly suggest that in times of severe finan-
cial crises that affect the real sector, potent “unconventional” monetary policy, which calm
financial markets, can be a promising strategy. However, both timing and intensity of policy
actions matter.

References
[1] Adrian, T., A. Moench, and H.S. Shin (2010), “Macro Risk Premium and Intermediary
Balance Sheet Quantities,” Federal Reserve Bank of New York, Staff Report No. 428

[2] Becker, S., L. Gruene, and W. Semmler (2007), “Comparing Second Order Approxima-
tion with Dynamic Programming,” Computational Economics, 30, 65–91.

[3] Bernanke, B., M. Gertler, and S. Gilchrist (1999), “The Financial Accelerator in a Quan-
titative Business Cycle Framework,” in J. Taylor and M. Woodford (eds), Handbook of
Macroeconomics, Amsterdam, North-Holland.

[4] Bruno, R. and R.W. Dimand (2009), “The Corridor of Stability in Tobin’s Keynesian
Model of Recession and Depression,” International Journal of Applied Economics &
Econometrics, XVII, 17–25.

[5] Brunnermeier, M. (2009), “Deciphering the Liquidity and Credit Crunch 2007–2008,”
Journal of Economic Perspectives, 23, 77–100.

[6] Brunnermeier, M. and Y. Sannikov (2010), “A Macroeconomic Model with the Financial
Sector,” manuscript, Dept. of Economics, Princeton University; see also their version
(2011).

32
 [7] Burnham, K.P. and D.R. Anderson (1998), Model Selection and Inference: A Practical
Information-Theoretical Approach New York: Springer–Verlag.

[8] Caballero, R. and A. Krishnamurtha (2008), “Global Imbalances and Financial

Fragility,” http://econ-www.mit.edu/files/366, MIT.

[9] Cardarelli, R., S. Elekdag, and S. Lall (2011), “Financial Stress and Economic Contrac-
tions,” Journal of Financial Stability, 7, 78–97.

[10] Christensen, I. and A. Dib (2008), “The Financial Accelerator in an Estimated New
Keynesian Model,” Review of Economic Dynamics, 11, 155–178.

[11] Cochrane, J. (2011), “Presidential Address: Discount Rates,” manuscript, University of

Chicago.

[12] Curdia, V. and M. Woodord (2009), “Credit Spreads and Monetary Policy,” Paper
presented at the Conference on Financial Market and Monetary Policy, Federal Reserve
Board, June 5–6.

[13] Davig, T. and C. Hakkio (2010), “What is the Effect of FInancial Stress on Economic
Activity?,” Economic Review, Federal Reserve Bank of Kansas City, Second Quarter,
35–62.

[14] Del Negro, M., G. Eggersson, A. Ferrero, and N. Kiyotaki (2011), “The Great Escape:
A Quantitative Evaluation of the Fed’s Non-Standard Policies,” Federal Reserve Bank
of New York Staff Report 520.

[15] Dimand, R. W. (2005), “Fisher, Keynes, and the Corridor of Stability,” American Jour-
nal of Economics and Sociology, 64, 185–199.

[16] Ehrmann, M., M. Ellison and N. Valla (2003), “Regime–dependent Impulse Response
Functions in a Markov-switching Vector Autoregressive Model,” Economic Letters, 78,
295–299.

[17] Ernst, E. and W. Semmler (2010), “Global Dynamics in a Model with Search and
Matching in Labor and Capital Markets,” Journal of Economic Dynamics and Control,
34, 1651–1679.

[18] Ernst, E. S. Mittnik, and W. Semmler (2010), “Interaction of Labor and Credit Mar-
ket Frictions: A Theoretical and Empirical Analysis,” paper presented at the Winter
Meeting of the Econometric Society, Atlanta, January 2010.

[19] Financial Stability Board (2009), “FSB Principles for Sound Compensation Practices—
Implementation Standards,” 25. September 2009.

[20] Geanakoplos, J. (2010), “The Leverage Cycle,” Cowles Foundation Paper No. 1304.

[21] Geanakoplos, J. and J. Farmer (2009), “The Virtues and Vices of Equilibrium and the
Future of Financial Economics,” Complexity 14, 11-38.

33
[22] Geneva Report on the World Economy 11 (2009), “Fundamental Principles of Financial
Regulations,” International Center for Monetary and Banking Studies.

[23] Gilchrist, S., A. Ortiz, and S. Zagrajsek (2009), “Credit Risk and the Macroeconomy:
Evidence from an Estimated DSGE Model,” manuscript, Department of Economics,
Boston University.

[24] Gorton, G.B. (2009), “Information, Liquidity, and the (Ongoing) Panic of 2007,” NBER
Working Paper No. 14649.

[25] Gorton, G.B. (2010), “Slapped by the Invisible Hand: The Panic of 2007,” Oxford:
Oxford University Press.

[26] Gruene, L. and W. Semmler (2004), “Using Dynamic Programming with Adaptive Grid
Scheme for Optimal Control Problems in Economics,” Journal of Economic Dynamics
and Control, 28, 2427–2456.

[27] Gruene, L. and W. Semmler (2005), “Default Risk, Asset Pricing, and Debt Control,”
Journal of Financial Econometrics 2005 3, 79–106.

[28] Gruene, L. and W. Semmler (2008), “Asset Pricing with Loss Aversion,” Journal of
Economic Dynamics and Control, 32, 3253–3274.

[29] Hakkio, C.S. and W.R. Keeton (2009), “Financial Stress: What Is It, How Can It Be
Measured, and Why Does It Matter?” Federal Reserve Bank of Kansas City, Economic
Review, Second Quarter, 5–50.

[30] Hall, E.R. (2010), Forward Looking Decision Making, Princeton: Princeton University
Press.

[31] Hamilton, J.D. (1989), “A New Approach to the Economic Analysis of Nonstationary
Time Series and the Business Cycle,” Econometrica 57, 357–384.

[32] Hansen, B.E. (2000), “Sample Splitting and Threshold Estimation,” Econometrica 68,
575–603.

[33] Hatzius, J., P. Hooper, F. Mishkin, K. Schoenholtz, and M.W. Watson (2010), Financial
Condition Indexes: A Fresh Look after the Financial Crisis,” NBER Working Paper No.
16150.

[34] He, Z. and A. Krishnamurthy (2008), “Intermediary Asset Pricing,” manuscript, Uni-
versity of Chicago.

[35] Hollo, D., M. Kremer and M. Lo Duca (2012), “CISS—A Composite Indicator of Sys-
temic Stress in the Financial System,” Working Paper, No. 1426, European Central
Bank.

[36] Hubrich, K. and R.J. Tetlow (2011), “Financial Stress and Economic Dynamics: The
Transmission of Crises,” unpublished working paper European Central Bank.

34
[37] Hsiao, C.-Y. and W. Semmler (2009), “Harmonic Fitting for Long Run Economic Data,
manuscript, www.newschool.edu/nssr/cem (there: "Financial Market")
[38] Illing, M. and Y. Liu (2006), “Measuring Financial Stress in a Developed Country: An
Application to Canada,” Journal of Financial Stability, 2, 243–265.
[39] Kindleberger, C. and R. Aliber (2005), Manias, Panics, and Crashes: A History of
Financial Crises, 5th ed., Wiley.
[40] Kliesen, K.L., M.T. Owyang and E.K. Vermann (2012), “Disentangling Diverse Mea-
sures: A Survey of Financial Stress Indexes,” Federal Reserve Bank of St. Louis Review,
September/October 2012, 94, 369–397.
[41] Koop, G., M. Pesaran and S. Potter (1996), “Impulse Responses in Nonlinear Multivari-
ate Models,” Journal of Econometrics, 74, 119–147.
[42] Mallick, S. K. and R. M. Sousa (2011), “The Real Effects of Financial Stress in the Euro
Zone,” NIPE Working Paper 12/2011, Department of Economics and Economic Policy
Research Unit (NIPE), University of Minho.
[43] Merton, R.C. (1974), “On the Pricing of Corporate Debt: The Risk Structure of the
Interest Rate,” Journal of Finance, 2, 449–470.
[44] Minsky, H.P. (1975), John Maynard Keynes, New York: Columbia University Press.
[45] Minsky, H.P. (1982), Can it Happen again?, Armonk: ME Sharpe.
[46] Mittnik, S. and W. Semmler (2012), “Regime Dependence of the Fiscal Multiplier,”
Journal of Economic Behavior & Organization, 83, 502–522.
[47] Mittnik, S. and P. Zadrozny (1993), “Asymptotic Distributions of Impulse Responses,
Step Responses, and Variance Decompositions of Estimated Linear Dynamic Models,”
Econometrica, 61, 857–887.
[48] Monnin, P. and T. Jokipii (2010), “The Impact of Banking Instability on the Real
Economy,” Working Paper 2010-5, Swiss National Bank.
[49] OECD (2012), OECD Monthly Economic Indicators.
[50] Reinhart. C. and K.S. Rogoff (2009), This Time Is Different: Eight Centuries of Finan-
cial Folly, Princeton: Princeton University Press.
[51] van Roye, B. (2012), “Financial Stress and Economic Activity in Germany,” unpublished
working paper, Kiel Institute for World Economy.
[52] Schinasi, G.J. (2005), “Preserving Financial Stability,” International Monetary Fund,
Economic Issues, 36.
[53] Semmler, W. and L. Bernard (2012), “Boom-Bust Cycles: Leveraging, Complex Securi-
ties, and Asset Prices,” Journal of Economic Behavior and Organization, 81, 442–465.

35
[54] Semmler, W., L. Gruene, and L. C. Oehrlein (2009), “Dynamic Consumption and Port-
folio Decisions with Time Varying Asset Returns,” Journal of Wealth Management, 12,
21–47.

[55] Tong, H. (1978), “On a Threshold Model,” in: Pattern Recognition and Signal Process-
ing, edited by C. H. Chen, Amsterdam: Kluwer.

[56] Tong, H. (1983), Threshold Models in Non-linear Time Series Analysis, New York:
Springer–Verlag.

[57] Tsay, R.S. (1998), “Testing and Modelling Multivariate Threshold Models,” Journal of
the American Statistical Association, 93, 1188–1202.

[58] Wong, C.S. and W.K. Li (1998), “A Note on the Corrected Akaike Information Criterion
for Threshold Autoregressive Models,” Journal of Time Series Analysis, 19, 113–124.

36
Appendix: Estimating Time Varying Credit Spreads

The calibration of the theoretical model in Section 3 is based on Hsiao and Semmler (2009)
and applies Fast Fourier Transformation (FFT) to German FSI as a proxy of the time–
varying default risk and credit cost. Estimating the periodic components by fitting linear
combinations of sine–cosine functions, i.e.,
n 
X 2π
2π

xt = ai sin (t − t0 ) + bi cos (t − t0 ) . (14)
i=1
τi τi

we obtain the coefficients reported in 6. The estimates in Tabel 6 are then used for the DP
solution. For details see Semmler et al. (2009). The close fit of (14) is illustrated in Figure
14.

Table 6: Coefficients of the harmonic fit of the real bond yield in (14)

i= 1 2 3 4 5 6
τi (month) 120 36 48 60 96 75
ai -0.263 0.0854 1.1592 0.1483 1.9493 0.1523
bi 0.0445 -0.0837 1.2705 0.6065 1.6699 1.2001

6 periods included
12

10

−2

−4

−6

−8
0 5 10 15 20 25 30

Figure 14: Actual and fitted FSI series for Germany

37